Re: [RFC] sched: Limit idle_balance() when it is being used too frequently

[Jason Low <jason.low2@hp.com>]

On Wed, 2013-07-17 at 09:25 +0200, Peter Zijlstra wrote:
> On Tue, Jul 16, 2013 at 03:48:01PM -0700, Jason Low wrote:
> > On Tue, 2013-07-16 at 22:20 +0200, Peter Zijlstra wrote:
> > > On Tue, Jul 16, 2013 at 12:21:03PM -0700, Jason Low wrote:
> > > > When running benchmarks on an 8 socket 80 core machine with a 3.10 kernel,
> > > > there can be a lot of contention in idle_balance() and related functions.
> > > > On many AIM7 workloads in which CPUs go idle very often and idle balance
> > > > gets called a lot, it is actually lowering performance.
> > > > 
> > > > Since idle balance often helps performance (when it is not overused), I
> > > > looked into trying to avoid attempting idle balance only when it is
> > > > occurring too frequently.
> > > > 
> > > > This RFC patch attempts to keep track of the approximate "average" time between
> > > > idle balance attempts per CPU. Each time the idle_balance() function is
> > > > invoked, it will compute the duration since the last idle_balance() for
> > > > the current CPU. The avg time between idle balance attempts is then updated
> > > > using a very similar method as how rq->avg_idle is computed. 
> > > > 
> > > > Once the average time between idle balance attempts drops below a certain
> > > > value (which in this patch is sysctl_sched_idle_balance_limit), idle_balance
> > > > for that CPU will be skipped. The average time between idle balances will
> > > > continue to be updated, even if it ends up getting skipped. The
> > > > initial/maximum average is set a lot higher though to make sure that the
> > > > avg doesn't fall below the threshold until the sample size is large and to
> > > > prevent the avg from being overestimated.
> > > 
> > > One of the things I've been talking about for a while now is how I'd
> > > like to use the idle guestimator used for cpuidle for newidle balance.
> > > 
> > > Basically based on the estimated idle time limit how far/wide you'll
> > > search for tasks to run.
> > > 
> > > You can remove the sysctl and auto-tune by measuring how long it takes
> > > on avg to do a newidle balance.
> > 
> > Hi Peter,
> > 
> > When you say how long it takes on avg to do a newidle balance, are you
> > referring to the avg time it takes for each call to CPU_NEWLY_IDLE
> > load_balance() to complete, or the avg time it takes for newidle balance
> > attempts within a domain to eventually successfully pull/move a task(s)?
> 
> Both :-), being as the completion time would be roughly equivalent for the
> top domain and the entire call.
> 
> So I suppose I was somewhat unclear :-) I initially started out with a
> simpler model, where you measure the avg time of the entire
> idle_balance() call and measure the avg idle time and compare the two.
> 
> Then I progressed to the more complex model where you measure the
> completion time of each domain in the for_each_domain() iteration of
> idle_balance() and compare that against the estimated idle time, bailing
> out of the domain iteration when the avg completion time exceeds the
> expected idle time.

Hi Peter,

For the more complex model, are you suggesting that each completion time
is the time it takes to complete 1 iteration of the for_each_domain()
loop? 

Based on some of the data I collected, a single iteration of the
for_each_domain() loop is almost always significantly lower than the
approximate CPU idle time, even in workloads where idle_balance is
lowering performance. The bigger issue is that it takes so many of these
attempts before idle_balance actually "worked" and pulls a tasks.

I initially was thinking about each "completion time" of an idle balance
as the sum total of the times of all iterations to complete until a task
is successfully pulled within each domain.

Jason